[GPU] onednn integration rls-v3.12-pc (#34446)

### Details:
```
commit 3e48e1f324199c61d935478e0787ebbc7e7616b6 (HEAD, origin/rls-v3.12-pc)
Author: Kealan Barbieri <kealan.barbieri@intel.com>
Date:   Mon Mar 2 11:20:01 2026 -0800
 
    xe: gemm: jit: fixup dp4a sums on xe3p
```

---------

Co-authored-by: Kim, Eddy <eddy.kim@intel.com>

Author: sungeunk

src/plugins/intel_gpu/include/intel_gpu/runtime/device_info.hpp

@@ -34,6 +34,9 @@ enum class gpu_arch {
     xe_hpc = 6,
     xe2 = 7,
     xe3 = 8,
+    xe3p_35_10 = 9,
+    xe3p_35_11 = 10,
+    xe3p_unknown = 11,
 };
 
 /// @brief Defines version of GFX IP

src/plugins/intel_gpu/src/graph/impls/ocl/kernels_cache.cpp

@@ -44,7 +44,7 @@
 #ifndef NOMINMAX
 # define NOMINMAX
 #endif
-#include "gpu/intel/microkernels/fuser.hpp"
+#include "gpu/intel/gemm/jit/include/gemmstone/microkernel/fuser.hpp"
 #endif
 
 namespace {
@@ -350,8 +350,7 @@ void kernels_cache::build_batch(const batch_program& batch, compiled_kernels& co
             std::vector<uint8_t> binary = kernels[0]->get_binary();
             kernels.clear();
             // Update binary and rebuild kernel
-            using namespace dnnl::impl::gpu::intel;
-            micro::fuseMicrokernels(binary, combined_source.c_str());
+            gemmstone::microkernel::fuse(binary, combined_source.c_str());
             _builder->build_kernels(binary.data(), binary.size(), KernelFormat::NATIVE_BIN, "", kernels);
 #else  // ENABLE_ONEDNN_FOR_GPU
             OPENVINO_THROW("[GPU] Can't compile kernel w/ microkernels as onednn is not available");

src/plugins/intel_gpu/src/graph/impls/ocl_v2/moe/moe_3gemm_gen_micro.cpp

@@ -223,7 +223,7 @@ void MoE3GemmMicroGenerator::init_microkernels(const kernel_impl_params& params,
     GPU_DEBUG_TRACE_DETAIL << "\t weight group size: " << group_size << "\n";
 
     micro::GEMMProblem problem_moe;
-    micro::GEMMProtocol::Options opts_moe;
+    micro::GEMMOptions opts_moe;
     opts_moe.slmPtr = true;
     opts_moe.kParallelLocal = !is_prefill;
     enum class MICRO_DIMENSIONALITY { NONE = -1, SCALAR = 0, VECTOR = 1, MATRIX = 2 };

src/plugins/intel_gpu/src/graph/impls/ocl_v2/moe/moe_gemm_gen_micro.cpp

@@ -149,7 +149,7 @@ void MoEGemmMicroGenerator::init_microkernels(const kernel_impl_params& params,
     GPU_DEBUG_TRACE_DETAIL << "init_microkernels for " << (is_prefill ? "prefill" : "generate") << " : Seq_len:" << n << " Ofm:" << m << " K:" << k << "\n";
 
     micro::GEMMProblem problem_moe;
-    micro::GEMMProtocol::Options opts_moe;
+    micro::GEMMOptions opts_moe;
     opts_moe.slmPtr = true;
     opts_moe.kParallelLocal = !is_prefill;
     enum class MICRO_DIMENSIONALITY { NONE = -1, SCALAR = 0, VECTOR = 1, MATRIX = 2 };

src/plugins/intel_gpu/src/graph/impls/ocl_v2/sdpa/sdpa_gen_micro.cpp

@@ -28,12 +28,8 @@ size_t get_subgroup_size(gpu_arch arch) {
     case gpu_arch::xe_hp:
     case gpu_arch::xe_hpg:
         return 8;
-    case gpu_arch::xe_hpc:
-    case gpu_arch::xe2:
-    case gpu_arch::xe3:
-        return 16;
     default:
-        return 0;
+        return 16;
     }
 }
 
@@ -1465,8 +1461,7 @@ void SDPAMicroGenerator::init_microkernels(const kernel_impl_params& params,
                                      is_paged_attention,
                                      is_prefill);
         break;
-    case gpu_arch::xe2:
-    case gpu_arch::xe3: {
+    default: {
         config = choose_config_xe2(static_cast<int32_t>(k_head_size),
                                    static_cast<int32_t>(nkeys_v),
                                    thin_q,
@@ -1476,8 +1471,6 @@ void SDPAMicroGenerator::init_microkernels(const kernel_impl_params& params,
                                    is_prefill);
         break;
     }
-    default:
-        break;
     }
 
     OPENVINO_ASSERT(config != nullptr);
@@ -1501,7 +1494,7 @@ void SDPAMicroGenerator::init_microkernels(const kernel_impl_params& params,
     problem_kq.A.layout = (is_paged_attention && !is_prefill) ? micro::MatrixLayout::N : micro::MatrixLayout::T;
 
     /* Set up microkernel options */
-    micro::GEMMProtocol::Options opts_kq;
+    micro::GEMMOptions opts_kq;
     opts_kq.localB = true;
     opts_kq.slmPtr = true;
 
@@ -1635,7 +1628,7 @@ void SDPAMicroGenerator::init_microkernels(const kernel_impl_params& params,
     }
 
     /* Set up microkernel options */
-    micro::GEMMProtocol::Options opts_vs;
+    micro::GEMMOptions opts_vs;
     opts_vs.localB = true;
     opts_vs.slmPtr = true;
 

src/plugins/intel_gpu/src/kernel_selector/kernel_selector_params.h

@@ -366,6 +366,9 @@ enum class gpu_arch {
     xe_hpc = 6,
     xe2 = 7,
     xe3 = 8,
+    xe3p_35_10 = 9,
+    xe3p_35_11 = 10,
+    xe3p_unknown = 11,
 };
 
 

src/plugins/intel_gpu/src/kernel_selector/micro_utils.hpp

@@ -17,27 +17,28 @@
 # define NOMINMAX
 #endif
 
-#include "gpu/intel/microkernels/package.hpp"
+#include "gpu/intel/gemm/jit/include/gemmstone/microkernel/package.hpp"
 #include "gpu/intel/gemm/jit/include/gemmstone/kernel_selector.hpp"
-#include "gpu/intel/gemm/jit/include/gemmstone/microkernel_provider.hpp"
-#include "gpu/intel/microkernels/shim.hpp"
+#include "gpu/intel/gemm/jit/include/gemmstone/microkernel_selector.hpp"
+#include "gpu/intel/gemm/jit/include/gemmstone/microkernel/shim.hpp"
 #include "common/utils.hpp"
 
 namespace micro {
 
-using Package = dnnl::impl::gpu::intel::micro::Package;
-using HWInformation = gemmstone::HWInformation;
+using Package = gemmstone::microkernel::Package;
+using HWInformation = gemmstone::microkernel::HWInformation;
 using GEMMProblem = gemmstone::GEMMProblem;
 using ABOffset = gemmstone::ABOffset;
 using GEMMStrategy = gemmstone::GEMMStrategy;
-using GEMMProtocol = dnnl::impl::gpu::intel::micro::GEMMProtocol;
+using GEMMProtocol = gemmstone::microkernel::Protocol;
+using GEMMOptions = gemmstone::microkernel::GEMMOptions;
 using MatrixLayout = gemmstone::MatrixLayout;
 using Type = gemmstone::Type;
 using SizeParams = gemmstone::SizeParams;
 using StrategyRequirement = gemmstone::StrategyRequirement;
-using ShimOptions = dnnl::impl::gpu::intel::micro::ShimOptions;
-using HostLanguage = dnnl::impl::gpu::intel::micro::HostLanguage;
-using Setting = dnnl::impl::gpu::intel::micro::Setting;
+using ShimOptions = gemmstone::microkernel::ShimOptions;
+using HostLanguage = gemmstone::microkernel::HostLanguage;
+using Setting = gemmstone::microkernel::Package::Setting;
 
 using dnnl::impl::utils::rnd_up_pow2;
 
@@ -72,18 +73,18 @@ struct MicroKernelPackage {
     }
 };
 
-inline Package select_gemm_microkernel(GEMMProtocol protocol, HWInformation hw_info, SizeParams sizes, const GEMMProblem &problem,
+inline Package select_gemm_microkernel(GEMMOptions &options, HWInformation hw_info, SizeParams sizes, const GEMMProblem &problem,
                                         const std::vector<StrategyRequirement> &reqs = std::vector<StrategyRequirement>(),
                                         void (*strategyAdjuster)(GEMMStrategy &strategy) = nullptr, gemmstone::SelectionObserver *observer = nullptr) {
-    return gemmstone::selectGEMMMicrokernel(protocol, hw_info, sizes, problem, reqs, strategyAdjuster, observer);
+    return gemmstone::microkernel::selectGEMM(options, hw_info, sizes, problem, reqs, strategyAdjuster);
 }
-inline Package select_gemm_microkernel(GEMMProtocol protocol, HWInformation hw_info, SizeParams sizes, const GEMMProblem &problem,
+inline Package select_gemm_microkernel(GEMMOptions &options, HWInformation hw_info, SizeParams sizes, const GEMMProblem &problem,
         gemmstone::SelectionObserver *observer) {
-    return gemmstone::selectGEMMMicrokernel(protocol, hw_info, sizes, problem, {}, nullptr, observer);
+    return gemmstone::microkernel::selectGEMM(options, hw_info, sizes, problem, {}, nullptr);
 }
 
 static inline int alignment_for_ld(int ld) {
-    return  gemmstone::alignmentForLD(ld);
+    return  gemmstone::microkernel::alignmentForLD(ld);
 }
 
 }  // namespace micro

src/plugins/intel_gpu/src/runtime/ocl/ocl_device.cpp

@@ -63,6 +63,9 @@ gpu_arch convert_ngen_arch(ngen::HW gpu_arch) {
         case ngen::HW::XeHPC: return gpu_arch::xe_hpc;
         case ngen::HW::Xe2: return gpu_arch::xe2;
         case ngen::HW::Xe3: return gpu_arch::xe3;
+        case ngen::HW::XE3P_35_10: return gpu_arch::xe3p_35_10;
+        case ngen::HW::XE3P_35_11: return gpu_arch::xe3p_35_11;
+        case ngen::HW::XE3P_UNKNOWN: return gpu_arch::xe3p_unknown;
         case ngen::HW::Gen10:
         case ngen::HW::Unknown: return gpu_arch::unknown;
     }

src/plugins/intel_gpu/thirdparty/CMakeLists.txt

@@ -14,8 +14,8 @@ if(ENABLE_ONEDNN_FOR_GPU)
         set(ONEDNN_BUILD_DIR "${CMAKE_CURRENT_BINARY_DIR}/onednn_gpu_build")
         set(ONEDNN_INSTALL_DIR "${CMAKE_CURRENT_BINARY_DIR}/onednn_gpu_install" CACHE PATH "Installation path for oneDNN GPU library")
         set(ONEDNN_PREFIX_DIR "${CMAKE_CURRENT_BINARY_DIR}/onednn_gpu_root")
-        set(ONEDNN_ENABLED_PRIMITIVES "CONCAT;CONVOLUTION;DECONVOLUTION;INNER_PRODUCT;MATMUL;REORDER;POOLING;REDUCTION;SDPA;RNN")
-        set(ONEDNN_ENABLED_ISA "XELP;XEHP;XEHPG;XEHPC;XE2;XE3")
+        set(ONEDNN_ENABLED_PRIMITIVES "CONCAT;CONVOLUTION;DECONVOLUTION;GATED_MLP;INNER_PRODUCT;MATMUL;REORDER;POOLING;REDUCTION;SDPA;RNN")
+        set(ONEDNN_ENABLED_ISA "ALL")
         set(DNNL_GPU_LIBRARY_NAME "openvino_onednn_gpu" CACHE STRING "Name of oneDNN library for Intel GPU Plugin")
 
         if(X86_64)